this was announced last week. The major changes are memory capacity and bandwidth. That means smaller batch sizes (i.e. inference) will be faster (1.4-1.7x) a

this was announced last week. The major changes are memory capacity and bandwidth. That means smaller batch sizes (i.e. inference) will be faster (1.4-1.7x) and larger batch sizes (i.e. training) will be possible. Current large batch sizes, however, won't be faster.